ggml : remove bit shuffling

[ggerganov]

Close #1241

• Drop Q4_2 support
• Changed bit-order for Q4 and Q5 (breaking change)
• Preplexity is perplexing as usual

---

New timings:

Model	Measure	F16	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0
7B	ms/tok @ 4th	128	50	54	75	83	75
7B	ms/tok @ 8th	123	44	52	53	58	72
13B	ms/tok @ 4th	332*	93	101	150	164	141
13B	ms/tok @ 8th	308*	81	96	96	104	136

• these numbers vary a lot since the model is on the 32GB limit of my MacBook

Old timings:

Model	Measure	F16	Q4_0	Q4_1	Q5_0	Q5_1	Q8_0
7B	ms/tok @ 4th	128	56	61	91	95	75
7B	ms/tok @ 8th	128	47	55	53	59	75
13B	ms/tok @ 4th	239	104	113	176	185	141
13B	ms/tok @ 8th	240	85	99	108	117	147

overall, all these numbers seem to have about +/- 10% variablility from run to run. not ideal benchmark, but not sure what else to do

[redthing1]

Is there a script to upgrade the old models to new? I don't have the source models because they're huge.

[sw]

Well this was a bit rushed, I think for such a significant change it would have been nice to allow a discussion.

Ah, that's unfortunate. I don't think there's a way to fix this efficiently for the ARM_NEON branch

Fix what? Were the NEON implementations not working in #1305 / #1384?

[ggerganov]

#1384 does not work for NEON because when we remove the vzip calls, the high bits are no longer in the right place.

This is the relevant section before this PR:

https://github.com/ggerganov/llama.cpp/blob/b608b55a3ea8e4760c617418538465449175bdb8/ggml.c#L3335-L3360

We were ORing the 5th bit after the vzip.
I didn't see a way to fix this without shuffling the tables or the bits in some way.

[M00N-MAN]

Is there a script to upgrade the old models to new? I don't have the source models because they're huge.

Hello All

Could someone please share the way of requantizing?

At least on this example https://huggingface.co/TheBloke/WizardLM-7B-uncensored-GGML/blob/main/README.md
What should be done with WizardLM-7B-uncensored.q5_1.bin or which tool need to by used for https://huggingface.co/ehartford/WizardLM-13B-Uncensored which is the source model.

upd:
./quantize --help
usage: ./quantize model-f32.bin [model-quant.bin] type [nthreads]
type = "q4_0" or 2
type = "q4_1" or 3
type = "q5_0" or 8
type = "q5_1" or 9
type = "q8_0" or 7

but seems man README.md has lack of !q4 variants and sheet matching -n value with quantization of selected model

[LostRuins]

Cool to see this is merged, i'm slightly confused though, seems like the converter is still writing the old file version (1) ?

[philpax]

Cool to see this is merged, i'm slightly confused though, seems like the converter is still writing the old file version (1) ?

I think that's intentional - the converter converts the f16 model to a GGJT v1 f16 model, which is then quantised to a GGJT v2 qX_Y model. (The saver always writes version 2, from what I can see.)

[LostRuins]

Doesn't seem like it though, I don't see references where File version 2 is set during quantization either.

Edit: i'm wrong. It's set in write_magic()

[TheBloke]

Could someone please share the way of requantizing? <img alt="image" width="594" src="https://user-example https://huggingface.co/TheBloke/WizardLM-7B-uncensored-GGML/blob/main/README.md What should be done with WizardLM-7B-uncensored.q5_1.bin or which tool need to by used for https://huggingface.co/ehartford/WizardLM-13B-Uncensored which is the source model.

Check my repos again. I've re-quantised all my GGMLs using the latest code, in q4_0, q5_0, q5_1 and q8_0 variants. So no need to do it yourself unless you want to.